Amid the continual rise in sophisticated image manipulations, developing an accurate image manipulation localization (IML) model has become indispensable. Precisely identifying and delineating tampered regions depends on a model’s ability to capture non-semantic discrepancies between forged and authentic content. Traditional CNN-based methods, however, falter at modeling long-range dependencies and subtle artifact traces. By contrast, Transformers—with their self-attention mechanism—excel at highlighting minute anomalies across the entire image. Furthermore, because tampered regions can range from imperceptibly small edits to large-scale splices, a truly effective detector must adapt dynamically across multiple scales—something fixed up- and down-sampling pipelines cannot achieve. To tackle these challenges, we introduce a Transformer-based architecture that (1) extracts and fuses multi-scale features, (2) incorporates neural architecture search cells tailored to each scale, and (3) employs edge-supervised learning to sharpen boundary detection. Extensive experiments on several public benchmarks demonstrate that our model outperforms existing approaches, charting a promising course for the future of media forensics.